Fine-Grained Image Retrieval

Fine-grained image retrieval searches for images that match specific details, such as a product attribute, a vehicle part, a medical finding, or a particular instance, rather than only a broad category. Common approaches learn visual embeddings with metric or contrastive learning, align text or sketches with images, and compare local regions or multiple image tokens; hashing and shortlist-and-rerank pipelines help scale search. Benchmarks span fashion and product search, vehicles, biological species, medical images, and composed queries that ask for some attributes to change while others remain. A central challenge is balancing precise local matching and robust open-vocabulary understanding with efficient indexing, especially when clutter, appearance changes, or rare categories make global similarity unreliable.